Face

Face Detection

DenseBox (2015)

Object detection/DenseBox

SSH (ICCV 2017)

SSH: Single Stage Headless Face Detector decrease inference time
scale-invariant
fixed size arcter to replace proposal

DSFD (CVPR 2019)

DSFD: Dual Shot Face Detector
Feature Enhance Module (FEM)

Face Feature Embedding

Could used for recognition with simple classifier

DDML (CVPR 2014)

Discriminative Deep Metric Learning for Face Verification in the Wild learn a nonlinear transformations and yield discriminative deep metric with a margin between positive and negative image pairs

Contrastive Loss

Same class => close
Different class => distance > margin m

FaceNet (CVPR 2015)

FaceNet: A Unified Embedding for Face Recognition and Clustering

Triplet Loss

inspirated by LMNN (large margin nearest neighbor)
Triplet Loss with VAE
Triplet selection/ present a novel online negative exemplar mining strategy which ensures consistently increasing difficulty of triplets as the network trains
disadv: data expansion when constituting the sample pairs

Centre Loss (ECCV 2016)

A Discriminative Feature Learning Approach for Deep Face Recognition add center Loss to softmax, hence the model discriminative power enhanced

\[L_c = \frac{1}{2} \sum^m_{i=1} {||{x_i - c_{y_i}}||}^2_2\]

../../_images/centre_loss.png

SphereFace (CVPR 2017)

SphereFace: Deep Hypersphere Embedding for Face Recognition in an angular space and penalises the angles between deep features
normalize W, optimize feature embedding and angle
disadv: requires a series of approximation in order to be computed,, resulted in unstable training. softmax loss used to stabilise training dominate the training process.
../../_images/SphereFace.png

CosFace (CVPR 2018)

add cosine margin penalty to the target logit admits easier implementation and relieves the need for joint supervision from the softmax loss